半监督极限学习机在化工过程软测量建模中的应用研究

发布时间：2018-03-02 07:26

本文关键词： 软测量半监督学习极限学习机交叉验证递推方法　出处：《浙江工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：化工过程软测量建模对于保障化工生产装置的连续平稳操作及提高其产品质量具有重要的作用。然而,半监督学习方法还很少应用于化工过程的软测量建模,导致大量无标签样本的信息无法有效利用。本文提出两种半监督软测量建模方法,主要研究内容和创新点如下:(1)针对目前大部分化工过程软测量模型没有利用无标签样本的缺点,提出将半监督极限学习机(Semi-supervised Extreme Learning Machine,SELM)应用到化工过程软测量建模中。SELM将无标签样本所携带的信息融合入软测量建模框架,有效提高了模型的预测精度及其可靠性。以重油催化裂化汽油收率软测量为例,与反向传播神经网络和极限学习机相比,验证了该方法的有效性。(2)SELM中惩罚参数?不能随有标签样本个数的变化自动变化,采用传统留一交叉验证方法的计算复杂度较大。提出了快速留一交叉验证半监督极限学习机(Fast-Leave-One-Out-Cross-Validation SELM,FSELM)方法,FSELM可以自动选取合理的惩罚参数?,而且避免了传统留一交差验证复杂度较大的问题。以田纳西-伊斯曼过程19种成分变量软测量为例,与主元回归和极限学习机相比较,表明FSELM在化工过程软测量建模预测中,预测效果更好。(3)FSELM中隐层节点个数需要靠经验选取,若选取不当,可能会导致模型过拟合。并且每次节点个数改变,都需要重新建模,计算效率较低。提出了节点递推更新快速留一交叉验证半督极限学习机(Recursive FSELM,RFSELM)方法,它能根据节点递增和递减模型的均方根误差,自动选择恰当的节点个数,在减小计算量的同时,提高了模型的预测可靠性。以填料塔液泛气速和重油催化裂化汽油收率在线预测为例,验证了该方法的有效性。
[Abstract]:Soft sensing modeling of chemical process plays an important role in ensuring the continuous smooth operation of chemical plant and improving the product quality. However, semi-supervised learning method is rarely used in soft sensor modeling of chemical process. As a result, a large number of unlabeled samples can not be used effectively. In this paper, two semi-supervised soft sensor modeling methods are proposed. The main research contents and innovations are as follows: (1) aiming at the shortcomings of most of the current soft sensor models in chemical processes that do not use unlabeled samples, Semi-supervised Extreme Learning machine (SELM) is applied to soft sensing modeling of chemical process. SELM fuses the information carried by unlabeled samples into soft sensor modeling framework. The prediction accuracy and reliability of the model are improved effectively. The effectiveness of the method is verified by comparing with the backpropagation neural network and the extreme learning machine in the case of soft sensing of heavy oil catalytic cracking gasoline yield, and the penalty parameters in SELM are verified. It can not change automatically with the change of the number of labeled samples, so it is more complicated to use the traditional method of retention and cross validation. A fast retention and one cross verification semi-supervised extreme learning machine (Fast-Leave-One-Out-Out-Out-Out-Out-Out-Out-Validation set FSELM) method is proposed to automatically select reasonable penalty parameters. In addition, it avoids the problem that the complexity of traditional residual difference verification is large. Taking 19 component variables of Tennessee Eastman process as an example, compared with principal component regression and extreme learning machine, it is shown that FSELM is used in modeling and prediction of soft sensing in chemical process. The number of hidden nodes in FSELM can be predicted better. The number of hidden nodes in FSELM needs to be selected by experience. If the selection is not proper, the model may be overfitted, and each time the number of nodes changes, the model needs to be remodeled. The computation efficiency is low. This paper presents a new method of node recursive renewal, which can automatically select the appropriate number of nodes according to the root-mean-square error of node increment and decrement model, and a fast cross-validated semi-supervised extreme learning machine called Recursive FSELMU RFSELM is proposed. At the same time, the reliability of the model is improved, and the effectiveness of this method is verified by the on-line prediction of gas flooding rate in packed column and the yield of heavy oil FCC gasoline.
【学位授予单位】：浙江工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TQ018

【相似文献】